arXiv subjects

Masateru Tsunoda

Publications and source records attributed to Masateru Tsunoda.

15 recordsLinked to original sources

arXiv2026

Detection of LLM-assisted Code Plagiarism Using k-gram Software Birthmarks

Large language models (LLMs) have significantly lowered the technical barrier to software plagiarism. By transforming existing source code while preserving its functionality, modern LLMs can generate semantically identical program that may evade traditional plagiarism detection techniques. Among such attacks, code paraphrasing modifies the syntax and structure of a program while preserving its behavior. This paper investigates whether software birthmarks can detect such LLM-assisted plagiarism. As a starting point, we employ k-gram software birthmarks based on unique k-grams of Java opcodes, with k ranging from 1 to 6. We employ three contemporary LLMs: ChatGPT-5.1-Codex-Mini, DeepSeek-V4-Flash, and Claude-Haiku-4.5. The dataset consists of individually compilable source files extracted from actively maintained BSD-2-Clause licensed Java projects. We further compare five similarity measures for birthmark matching: cosine similarity, Dice index, Jaccard coefficient, Simpson index, and edit-distance-based similarity. The results demonstrate that k-gram software birthmarks remain effective for detecting LLM-assisted plagiarism. Among the evaluated models, ChatGPT-5.1-Codex-Mini generated the most difficult-to-detect clones. Furthermore, the findings confirm the higher performance of coding-oriented models for plagiarism task.

cs.SE↗

arXiv2026

Project-wise Comparison of Software Birthmarks Using Weighted Partial Similarity

Software birthmarks provide a robust approach to detecting code plagiarism even under substantial modifications, while distinguishing independently developed software. Existing similarity measures are typically applied at the module level (e.g., source or class files). However, in practice, software reuse often occurs at the project level, where only a subset of modules may be reused. This setting introduces two key challenges: (1) partial reuse, where reused modules constitute only a small fraction of the project, and (2) incidental similarity from small modules, which can lead to false positives. In this paper, we establish a framework for project-wise birthmark comparison based on a symmetric aggregation of module-level similarities. On top of this framework, we propose two complementary mechanisms to address the above challenges. First, we introduce a weighting scheme that assigns higher importance to larger modules, reducing the influence of noisy matches from small modules. Second, we propose a partial similarity method that focuses on the top fraction of highly similar module pairs, enabling robust detection of partial reuse. We evaluate the proposed approach on 35 open-source Java projects across ten categories, where different versions of the same project are treated as reuse cases. The dataset and experimental artifacts are made publicly available to support reproducibility. Performance is assessed using two complementary properties of software birthmarks, resilience and credibility, combined via their harmonic mean. The results show that the proposed method consistently outperforms existing approaches, achieving robust and stable detection of partial code reuse at the project level.

cs.SE↗

arXiv2026

A feasibility study on filtering low-accessibility web pages considering color vision deficiency

Recently, the importance of universal design has increased. Color universal design (CUD) is one type of universal design that takes people with color vision deficiency (CVD) into consideration. Websites are important media for providing various types of information and functions. Therefore, it is essential to enhance the accessibility of web pages by incorporating CUD principles. The goal of our study is to help improve the accessibility of web pages. Our approach is to automatically filter low-accessibility web pages. To evaluate the feasibility of this approach, we conducted an experiment using 21 web pages. The prediction model identified low-accessibility pages with reasonable accuracy, achieving a maximum AUC of 0.76.

cs.IR↗

arXiv2024

Personalization of Code Readability Evaluation Based on LLM Using Collaborative Filtering

Code readability is an important indicator of software maintenance as it can significantly impact maintenance efforts. Recently, LLM (large language models) have been utilized for code readability evaluation. However, readability evaluation differs among developers, so personalization of the evaluation by LLM is needed. This study proposes a method which calibrates the evaluation, using collaborative filtering. Our preliminary analysis suggested that the method effectively enhances the accuracy of the readability evaluation using LLMs.

cs.SE↗

arXiv2024

Influence of Skill and Knowledge of Programmers on Program Behavior Visualization by CS Unplugged

Computer science unplugged (CS unplugged) is a method of teaching computer science and computational thinking. It does not use a computer but employs physical materials. As CS unplugged, past studies proposed a new method to visualize programming behavior, and evaluated the method based on the understanding test of program. However, the past studied did not consider the factors affecting the test such as participants knowledge. This study analyzed the relationship between the factors and the test results.

cs.CY↗

arXiv2024

The Impact of Defect (Re) Prediction on Software Testing

Cross-project defect prediction (CPDP) aims to use data from external projects as historical data may not be available from the same project. In CPDP, deciding on a particular historical project to build a training model can be difficult. To help with this decision, a Bandit Algorithm (BA) based approach has been proposed in prior research to select the most suitable learning project. However, this BA method could lead to the selection of unsuitable data during the early iteration of BA (i.e., early stage of software testing). Selecting an unsuitable model can reduce the prediction accuracy, leading to potential defect overlooking. This study aims to improve the BA method to reduce defects overlooking, especially during the early testing stages. Once all modules have been tested, modules tested in the early stage are re-predicted, and some modules are retested based on the re-prediction. To assess the impact of re-prediction and retesting, we applied five kinds of BA methods, using 8, 16, and 32 OSS projects as learning data. The results show that the newly proposed approach steadily reduced the probability of defect overlooking without degradation of prediction accuracy.

cs.SE↗

arXiv2024

An Empirical Study of the Impact of Test Strategies on Online Optimization for Ensemble-Learning Defect Prediction

Ensemble learning methods have been used to enhance the reliability of defect prediction models. However, there is an inconclusive stability of a single method attaining the highest accuracy among various software projects. This work aims to improve the performance of ensemble-learning defect prediction among such projects by helping select the highest accuracy ensemble methods. We employ bandit algorithms (BA), an online optimization method, to select the highest-accuracy ensemble method. Each software module is tested sequentially, and bandit algorithms utilize the test outcomes of the modules to evaluate the performance of the ensemble learning methods. The test strategy followed might impact the testing effort and prediction accuracy when applying online optimization. Hence, we analyzed the test order's influence on BA's performance. In our experiment, we used six popular defect prediction datasets, four ensemble learning methods such as bagging, and three test strategies such as testing positive-prediction modules first (PF). Our results show that when BA is applied with PF, the prediction accuracy improved on average, and the number of found defects increased by 7% on a minimum of five out of six datasets (although with a slight increase in the testing effort by about 4% from ordinal ensemble learning). Hence, BA with PF strategy is the most effective to attain the highest prediction accuracy using ensemble methods on various projects.

cs.SE↗

arXiv2024

On Applying Bandit Algorithm to Fault Localization Techniques

Developers must select a high-performance fault localization (FL) technique from available ones. A conventional approach is to try to select only one FL technique that is expected to attain high performance before debugging activity. In contrast, we propose a new approach that dynamically selects better FL techniques during debugging activity.

cs.SE↗

arXiv2024

Building Defect Prediction Models by Online Learning Considering Defect Overlooking

Building defect prediction models based on online learning can enhance prediction accuracy. It continuously rebuilds a new prediction model, when a new data point is added. However, a module predicted as "non-defective" can result in fewer test cases for such modules. Thus, a defective module can be overlooked during testing. The erroneous test results are used as learning data by online learning, which could negatively affect prediction accuracy. To suppress the negative influence, we propose to apply a method that fixes the prediction as positive during the initial stage of online learning. Additionally, we improved the method to consider the probability of the overlooking. In our experiment, we demonstrate this negative influence on prediction accuracy, and the effectiveness of our approach. The results show that our approach did not negatively affect AUC but significantly improved recall.

cs.SE↗

arXiv2023

Selecting Source Code Generation Tools Based on Bandit Algorithms

Background: Recently, code generation tools such as ChatGPT have drawn attention to their performance. Generally, a prior analysis of their performance is needed to select new code-generation tools from a list of candidates. Without such analysis, there is a higher risk of selecting an ineffective tool, negatively affecting software development productivity. Additionally, conducting prior analysis of new code generation tools takes time and effort. Aim: To use a new code generation tool without prior analysis but with low risk, we propose to evaluate the new tools during software development (i.e., online optimization). Method: We apply the bandit algorithm (BA) approach to help select the best code-generation tool among candidates. Developers evaluate whether the result of the tool is correct or not. When code generation and evaluation are repeated, the evaluation results are saved. We utilize the stored evaluation results to select the best tool based on the BA approach. Our preliminary analysis evaluated five code generation tools with 164 code generation cases using BA. Result: The BA approach selected ChatGPT as the best tool as the evaluation proceeded, and during the evaluation, the average accuracy by the BA approach outperformed the second-best performing tool. Our results reveal the feasibility and effectiveness of BA in assisting the selection of best-performing code generation tools.

cs.SE↗

arXiv2023

Software Defect Prediction by Online Learning Considering Defect Overlooking

Building defect prediction models based on online learning can enhance prediction accuracy. It continuously rebuilds a new prediction model when adding a new data point. However, predicting a module as "non-defective" (i.e., negative prediction) can result in fewer test cases for such modules. Therefore, defects can be overlooked during testing, even when the module is defective. The erroneous test results are used as learning data by online learning, which could negatively affect prediction accuracy. In our experiment, we demonstrate this negative influence on prediction accuracy.

cs.SE↗

arXiv2022

Relationship between Gender and Code Reading Speed in Software Development

Recently, workforce shortage has become a popular issue in information technology (IT). One solution to increasing the workforce supply is to increase the number of female IT professionals. This is because there is gender imbalance in information technology area. To accomplish this, it is important to suppress the influence of biases, such as the belief that men are more suited for careers in science and technology than women, and to increase the choice of careers available to female professionals. To help suppress the influence of gender bias, we analyzed the relationship between gender and code reading speed in the field of software development. Certain source codes require developers to use substantial memory to properly understand them, such as those with many variables that frequently change values. Several studies have indicated that the performance of memory differs in males and females. To test the veracity of this claim, we analyzed the influence of gender on code-reading speed through an experiment. Pursuant to this, we prepared four programs that required varied amounts of memory to properly understand them. Then, we measured the time required by each of the 17 male and 16 female subjects (33 subjects in total) to comprehend the different programs. The results suggest that there is no explicit difference between male and female subjects in this regard, even in the case of programs that require high memory capacities for proper understanding.

cs.SE↗

arXiv2022

Analyzing and Calibrating Risk Assessment by Software Developers

In software project management, risk management is a critical factor. Project managers use existing lists of risk or perform brainstorming to identify the risks. However, it is not easy to perceive all the risks objectively. As a result, some risks are perceived based on subjective impression, which leads to risk biases. So, our goals are (i) We clarify the risk perception of developers to enhance the reliability of the brainstorming, and (ii) we calibrate the risk assessment based on a mathematical model to make more accurate risk list. In the analysis, we collected data concerning the risk perception of 69 professional software developers via a questionnaire. The average number of years of experience among these professionals was 18.3. Using the dataset, we applied factor analysis to clarify the factors that affect the evaluation of risk impact. The questionnaire was based on the risk perception theory established by Slovic, in which "dread" and "unknown" are the major factor of risk perception. The analysis result shows that (i) risk experience (i.e., whether a developer actually faced the risk or not) sometimes affects risk assessment (evaluation of risk impact), (ii) risk perception is considered to be based on unknown and dread factors, and (iii) risk assessment can be calibrated by a mathematical model (the average absolute error was 0.20).

cs.SE↗

arXiv2022

Measuring Cognitive Load of Software Developers Based on Nasal Skin Temperature

It has recently become increasingly important to measure the cognitive load of developers. This is because continuing with a development under a high cognitive load may cause human errors. Therefore, to automatically measure the cognitive load, existing studies have used biometric measures such as brain waves and the heart rate. However, developers are often required to equip certain devices when measuring them, and can therefore be physically burdened. In this study, we evaluated the feasibility of non-invasive biometric measures based on the nasal skin temperature. The nasal skin temperature has been widely used in other fields to measure mental status. In the present experiment, the subjects created small Java programs, and we estimated their cognitive load using the proposed biometric measures based on the nasal skin temperature. As a result, the proposed biometric measures were shown to be more effective than non-biometric measures. Hence, biometric measures based on nasal skin temperature are promising for estimating the cognitive load of developers.

cs.SE↗

arXiv2020

A Simulation Study of Bandit Algorithms to Address External Validity of Software Fault Prediction

Various software fault prediction models and techniques for building algorithms have been proposed. Many studies have compared and evaluated them to identify the most effective ones. However, in most cases, such models and techniques do not have the best performance on every dataset. This is because there is diversity of software development datasets, and therefore, there is a risk that the selected model or technique shows bad performance on a certain dataset. To avoid selecting a low accuracy model, we apply bandit algorithms to predict faults. Consider a case where player has 100 coins to bet on several slot machines. Ordinary usage of software fault prediction is analogous to the player betting all 100 coins in one slot machine. In contrast, bandit algorithms bet one coin on each machine (i.e., use prediction models) step-by-step to seek the best machine. In the experiment, we developed an artificial dataset that includes 100 modules, 15 of which include faults. Then, we developed various artificial fault prediction models and selected them dynamically using bandit algorithms. The Thomson sampling algorithm showed the best or second-best prediction performance compared with using only one prediction model.

cs.SE↗